How Anthropic Scales Agents with Orchestrator-Worker Patterns
TRIGGER
Complex research queries require exploring multiple independent directions simultaneously, but single-agent systems hit context window limits and execute searches sequentially, making them too slow and limited for breadth-first exploration.
APPROACH
Anthropic built a multi-agent architecture where a lead agent (Claude Opus 4) analyzes queries, develops strategy, and spawns specialized subagents (Claude Sonnet 4) that operate in parallel with separate context windows. Input: user research query. Output: synthesized answer with sources. The lead agent decomposes queries into subtasks with explicit objectives, output formats, tool guidance, and task boundaries for each subagent. Subagents act as intelligent filters—searching, gathering, and condensing information before returning compressed results to the lead agent for final synthesis. Results: multi-agent system outperformed single-agent Claude Opus 4 by 90.2% on internal research eval; parallel execution cut research time by up to 90% for complex queries.
PATTERN
“Each subagent gets its own full context window—that's the scaling mechanism, not parallelism. Multi-agent research outperforms single-agent because reasoning capacity multiplies with agent count, not just speed.”
✓ WORKS WHEN
- Task involves pursuing 3+ independent directions that don't need to share context
- Information to process exceeds single context window limits (~200k tokens)
- Value of task outcome justifies 15× token cost compared to single-agent chat
- Queries are breadth-first (e.g., 'find all board members of S&P 500 IT companies') rather than depth-first
- Subagent results can be meaningfully compressed before passing to coordinator
✗ FAILS WHEN
- Tasks require all agents to share the same context or have many dependencies between steps
- Real-time coordination between agents is needed (agents aren't yet great at delegating in real time)
- Task is primarily sequential with each step depending on previous results (most coding tasks)
- Query is simple fact-finding that one agent with 3-10 tool calls can handle
- Token economics don't support 15× cost multiplier for the task value